针对X射线安检场景中违禁品目标检测精度低,检测模型过于复杂的问题,在YOLOv7-Tiny模型的基础上,提出了一种新的轻量化检测方法.首先在骨干网络中融合改进的轻量化模块GhostNetV2,在减少模型参数的同时,提高训练效率;其次在YOLOv7-Tiny的颈部网络部分加入金字塔拆分注意力机制,有效解决参数减少导致的提取特征不足问题,提高背景复杂以及多尺度目标回归的准确性;最后,通过使用归一化 Wasserstein距离方法来度量损失,替代了原有的Intersection over Union度量,降低了小目标位置偏差的敏感性,增强了小目标的回归准确性.实验结果表明,改进模型在SIXray、CLCXray和OPIXray数据集上平均检测精度达到92.9％、76.2％和91.2％,相比原始算法分别提升了 6.5％、2％和1.8％;所提出模型在轻量化的同时能够进一步提高检测能力,可以满足实时检测要求,具有较好的应用价值.
X-ray Security Contraband Detection on Improved YOLOv7-Tiny
In response to the issues of low accuracy and poor real-time performance in the X-ray security inspection scenario,a new lightweight detection method has been proposed based on the YOLOv7-Tiny model.Firstly,a lightweight GhostNetV2 module was introduced into the backbone network to reduce model parameters while improving training efficiency.Secondly,a Pyramid Split Attention mechanism was incorporated into the neck network of YOLOv7-Tiny to effectively address the problem of insufficient feature extraction due to parameter reduction,enhancing accuracy in detecting complex backgrounds and multi-scale objects.Finally,the Normalized Wasserstein Distance method was used to measure the loss,replacing the original Intersection over Union metric,which reduces sensitivity to small object position deviations and enhances accuracy in regressing small objects.Experimental results show that the improved model achieves an average detection accuracy of 92.9％,76.2％,and 91.2％on the SIXray,CLCXray,and OPIXray datasets,representing 6.5％,2％,and 1.8％improvement over the original algorithm,respectively.The proposed model,while being lightweight,further enhances detection capabilities and can meet real-time detection requirements,demonstrating significant practical value.
deep learningcontraband testingfeature extractionmodel lightweigh
万方数据
维普
